Re: [RFC PATCH net-next v2 1/2] tcp: Add net.ipv4.tcp_purge_receive_queue sysctl

On 15/7/26 23:15, Eric Dumazet wrote:

On Wed, Jul 15, 2026 at 4:54 PM Leon Hwang [off-list ref] wrote:

[...]

My thoughts are:

out_of_order_queue has been forgotten. skbs could be there and still
'block devmem'

Yes. The tp->out_of_order_queue should also be purged here so that
out-of-order SKBs cannot continue holding devmem.

WRITE_ONCE(tp->copied_seq, tp->rcv_nxt) is certainly wrong, because
read() will return 0, instead of -1 (errno = EPIPE or ECONNRESET)
So the application will not know a RST was received :/

Thanks for pointing this out. When sysctl_tcp_purge_receive_queue is
enabled, read() must report the reset error.

The reason read() returns 0 is that SOCK_DONE, which was set when the
FIN was processed, is checked before sk_err in tcp_recvmsg_locked(). I
think clearing SOCK_DONE after purging the queues could let read()
observes the error installed by tcp_done_with_error().

Will also update the packetdrill test to expect -1/EPIPE.

I think that BSD and linux implementations have historically retained
acknowledged,
buffered receive data upon RST to allow applications to drain data
already ACKed prior to the reset.

Agreed. The new sysctl is disabled by default specifically to preserve
this existing behavior.

Enabling it is an explicit opt-in to discard buffered receive data,
release the associated resources promptly, and report the reset error
immediately.

Will update the change in ip-sysctl.rst documentation with this
application-visible tradeoff.

Adding a narrow sysctl specifically for CLOSE_WAIT creates
inconsistent behavior across TCP states.

My bad. After reading the RFC 9293 section 3.10.7.4. [1] again, "All
segment queues should be flushed." should apply to these states:

      -  ESTABLISHED STATE
      -  FIN-WAIT-1 STATE
      -  FIN-WAIT-2 STATE
      -  CLOSE-WAIT STATE

So, the sysctl-controlled purge should be consistently applied to these
four states.

[1] https://www.rfc-editor.org/rfc/rfc9293.html#section-3.10.7.4

Thanks,
Leon